Run an explicitly authorized overnight collaboration with one desktop portal, a durable CLI writer and independent CLI review. Use for unattended work, not routine short tasks.
The English summary is being prepared. Showing the original description.
See how to installReview the instructions the agent will receive before you install.
Default human portal: Codex Desktop. Default writer: Claude Code CLI; reviewer: fresh independent Codex CLI. Verify actual available routes, not model-picker labels. Other providers or bb require explicit selection. A fresh review does not authorize launching a new agent; use only the roster allowed in the contract. Read mesh-comms-core and the relevant desktop adapter. Use sprint-overseer or gate-keeper only as the agreed risk level requires. The desktop is not a second writer.
Read the canonical Overnight Cookbook, especially chapters 2, 4-12, 16, 18-20 and Appendices A, I and J. These are the maintained procedures; historical Part A/B/C and M-number references are not separate authorities.
Confirm back the resolved run contract before starting: exact project/branch, human goal, allowed writes, forbidden effects, named writer and independent reviewer, model and harness routes, budget/time limits, first useful delivery, expected unattended period, alert owner/SLA, and local versus external acceptance. A skill is not a launch or spending authorization. Preserve dirty/staged work and identify single-writer leases.
Prove a real first round trip with each critical-path peer and one harmless tool action. Render, task re-entry, nonce ACK and completed work are separate evidence. Qualify a next-turn mechanism after arming; a process or scheduled-task ID is not proof it wakes the model. Keep a lifecycle sentinel responsible until joint closeout. Reconcile the whole current outstanding request set before moving a cursor; a shallow ledger tail cannot prove no work is pending.
Before unattended reliance, exercise the fixture fault matrix in the liveness runbook. Use structured, generation-fenced heartbeat records; distinguish WAKE_SEEN from HEALTH_VALIDATED. Test credentials and the actual human wake ingress before absence. Supervision must survive the agent's credential failure. Declare shared-host failure domains: use an independent observer or record the human's explicit risk acceptance. A substitute verifier does not cancel the duty to alert the human to a dead gatekeeper. If a required mechanism is unavailable, report UNPROVEN/BLOCKED; do not invent it.
Read own inbox, reconcile scoped outstanding nonces and artifacts, then classify progress, liveness and health separately. A missed ACK is an investigation trigger, not death or permission to replace a peer. Mirror material direct sends to the ledger and require receiver ACK; uncertainty requires inspection before retry. See desktop-live-comms for desktop lanes and mesh-comms-core for CLI lanes.
Deliver useful increments during the first hour. Use bounded review/repair rounds with clear acceptance and escalation; do not let process work substitute for product work. Only release the next specifically authorized milestone after an independent reviewer reruns relevant proof. The runner recommends; it cannot mint its own gate approval. Relayed human instructions need quoted/file-backed evidence and must not conflict with the latest direct request. Maintain checkpoints, decision/change records, hashes, and the automation registry. No secret values in evidence or comms.
For a Claude Code wedge, use the cookbook's qualified signature and recovery ladder (Appendix I.4), not spinner appearance alone. Preserve CPU/children/RSS, checkpoint, staged state and open nonces. At most one controlled Escape after the full signature, then verify postconditions and record RESUME. Diagnose recurrence; no blind restart. Historical watchdog environment knobs require runtime support checks, not assumption. Run heavy verification as bounded commands with retained output. Track stall/detection latency and useful progress, not just wake counts.
Stage 1 is local implementation plus independent review/tests and integrity checks. Stage 2 is any required credentialed, external, web, or human acceptance. Never claim Stage 1 proves Stage 2. Complete the final handoff, required peer closeout ACKs and human report before retiring the task-owned supervision; record what remains active. No push, merge, deployment, production mutation or new agent without the specific authority in the contract. Never use Ctrl-C as routine Codex clearing.
Report: topology; exact roles/task IDs; artifact hashes; proof commands/results; Stage-1/Stage-2 status; liveness evidence; unresolved blockers; next human decision.
No reviews yet. Tell others what you thought.
No login needed. Share how it worked for you, ask a question, or note where it did not run.
No comments yet. Be the first to write one.
Method for adding agent comments and work to a shared file without overlapping other agents. Creates isolated workspaces per agent with UTC timestamps. Use when told to add comments, reviews, or work to any file.
The English summary is being prepared. Showing the original description.
Check any artifact against an explicit specification and report adherence using the add-comments workspace format. Use when validating plans, scripts, docs, or code changes against a source-of-truth spec.
The English summary is being prepared. Showing the original description.
Join the Antigravity CLI (agy, Gemini 3.5 Flash) to the Interlateral mesh as a native CLI peer with its own tmux session, identity stamping, and the agy.js send helper. Use this instead of the Antigravity desktop-app CDP path.
The English summary is being prepared. Showing the original description.
Agents work in parallel on the same task; judges evaluate and select the winner.
The English summary is being prepared. Showing the original description.
Boot the explicitly selected CLI(s) as the human's concierge: a project-aware side channel that sits ABOVE the running agent system in a shared writable terminal. It answers status questions and handles the human's ad hoc tasking while the working agents collaborate on the mesh — it is not a worker, manager, or gatekeeper, and team progress never waits for it.
The English summary is being prepared. Showing the original description.
Multiple agents draft a structured document through federated co-authorship and formal voting.
The English summary is being prepared. Showing the original description.